Review and Motivation

Transcription

1 Review and Motivation We can model and visualize multimodal datasets by using multiple unimodal (Gaussian-like) clusters. K-means gives us a way of partitioning points into N clusters. Once we know which points go to which cluster, we can estimate a Gaussian mean and covariance for that cluster. We have introduced the idea of writing what you want to do as a function to be optimized (maximized or minimized). Maximum likelihood estimation to fit mean and covariance parameters of a Gaussian is a good example of this.

2 Review and Motivation want to do MLE of mixture of Gaussian parameters But this is hard, because of the summation in the mixture of Gaussian equation (can t take the log of a sum). If we knew which point contribute to which Gaussian component, the problem would be a lot easier (we can rewrite so that the summation goes away) So... let s guess which point goes with which component, and proceed with the estimation. We were unlikely to guess right the first time, but based on our initial estimation of parameters, we can now make a better guess at pairing points with components. Iterate This is the basic idea underlying the EM algorithm.

3 EM Algorithm What makes this estimation problem hard? 1) It is a mixture, so log-likelihood is messy 2) We don t directly see what the underlying process is

4 EM Algorithm What is the underlying process? procedure to generate a mixture of gaussians for i=1:n end generate a uniform U(0,1) random number to determine which of K components to draw a sample from (based on probabilities pi_k generate a sample from a Gaussian N(mu_k, Sigma_k)

5 EM Algorithm equivalent procedure to generate a mixture of gaussians for k=1:k end compute number of samples n_k = round(n * pi_k) to draw from the k-th component Gaussian generate n_k samples from Gaussian N(mu_k, Sigma_k) + + =

6 EM Algorithm labels Suppose some oracle told us which point comes from which Gaussian. How? By providing a latent variable z_nk which is 1 if point n comes from the kth component Gaussian, and 0 otherwise (a 1 of K representation)

7 EM Algorithm labels This lets us recover the underlying generating process decomposition: = + +

8 EM Algorithm labels And we can easily estimate each Gaussian, along with the mixture weights! N1 points N2 points N3 points = + + estimate pi_k = Nk / N estimate mu_1, Sigma_1 estimate mu_2, Sigma_2 estimate mu_3, Sigma_3

9 EM Algorithm Remember that this was a problem... how can I make that inner sum be a product instead???

10 EM Algorithm Remember that this was a problem... how can I make that inner sum be a product instead??? Again, if an oracle gave us the values of the latent variables (component that generated each point) we could work with the complete log likelihood and the log of that looks much better!

11 EM Algorithm Remember that this was a problem... how can I make that inner sum be a product instead??? Again, if an oracle gave us the values of the latent variables (component that generated each point) we could work with the complete log likelihood and the log of that looks much better!

12 Latent Variable View note: for a given n, there are k of these latent variables, and only ONE of them is 1 (all the rest are 0)

13 Latent Variable View This is thus equivalent to note: for a given n, there are k of these latent variables, and only ONE of them is 1 (all the rest are 0)

14 Latent Variable View

15 Latent Variable View can be estimated separately can be estimated separately can be estimated separately

16 Latent Variable View can be estimated separately can be estimated separately can be estimated separately these are coupled because the mixing weights all sum to 1, but it is no big deal to solve

17 EM Algorithm Unfortunately, oracle s don t exist (or if they do, they don t want to talk to us) So we don t know values of the the z_nk variables What EM proposes to do: 1) compute p(z X,theta), the posterior distribution over z_nk, given our current best guess at the values of theta 2) compute the expected value of the log likelihood ln(p(x,z theta)) with respect to the distribution p(z X,theta) 3) find theta_new that maximizes that function. This is our new best guess at the values of theta. 4) iterate...

18 Insight Since we don t know the latent variables, we instead take the expected value of the log likelihood with resepect to their distribution. In the GMM case, this is equivalent to softening the binary latent variables to continuous ones (the expected values of the latent variables) unknown discrete value 0 or 1 known continuous value between 0 and 1

19 Insight So now, after replacing the binary latent variables with their continuous expected values: all points contribute to the estimation of all components each point has unit mass to contribute, but splits it across the K components the amount of weight a point contributes to a component is proportional to the relative likelihood that the point was generated by that component

20 Latent Variable View (with oracle) can be estimated separately can be estimated separately can be estimated separately these are coupled because the mixing weights all sum to 1, but it is no big deal to solve

21 Latent Variable View (with EM, ) at iteration i i a constant i i i i can be estimated separately can be estimated separately i i can be estimated separately these are coupled because the mixing weights all sum to 1, but it is no big deal to solve

22 EM Algorithm for GMM E ownership weights M means covariances mixing probabilities

23 Gaussian Mixture Example: Start Copyright 2001, Andrew W. Moore

24 After first iteration

25 After 2nd iteration

26 After 3rd iteration

27 After 4th iteration

28 After 5th iteration

29 After 6th iteration

30 After 20th iteration

31 Recall: Labeled vs Unlabeled Data labeled Easy to estimate params (do each color separately) unlabeled Hard to estimate params (we need to assign colors)

32 EM produces a Soft labeling each point makes a weighted contribution to the estimation of ALL components

33 General EM Evaluate

34 Intuitive Explanation (in terms of function maximization and lower bounds) f(x) x1

35 Intuitive Explanation f(x) b1(x) x1

36 Intuitive Explanation f(x) b1(x) x1 x2

37 Intuitive Explanation f(x) b2(x) b1(x) x1 x2

38 Intuitive Explanation f(x) Why does this work? By construction, b1(x1) = f(x1) b1(x2) >= b1(x1) [it is a maximum] b1(x) b2(x) f (x2) >= b1(x2) [b1 is a lower bound] so, it is guaranteed that f(x2) >= f(x1) and in general, at each iteration f(xnew) >= f(xold) x1 x2 If f(x) is bounded above, then process should converge to a (local) maximum

39 More Rigorous Proof We will use Jensen s inequality for convex functions (see, for example, Bishop, PRML, p 56) With some manipulation, and reversing the inequality because log is a concave rather than convex function...

40 Proof that EM works this is function f(x) in our earlier picture definition of probability. Now use Jensen s inequality... lower bound b(x) in our earlier picture. Bishop calls this L(q,\theta)

41 Proof that EM works note, that when theta = theta_old if we expand this equation L(q,\theta_old) becomes just Showing that our lower bound touches the function ln p(x theta) at the current estimate theta_old (as promised by our earlier picture!)

42 Proof that EM works

43 Proof that EM works

44 Proof that EM works So... and we have increased our log likelihood. Therefore, finding argmax of L(theta,theta_old) is a good thing to do. But wait, there s more...

45 Proof that EM works This is the expected value Q(theta,theta_old) computed in the E-step of EM!!!! this doesn t depend on \theta. ignore it.

46 Proof that EM works This is the expected value Q(theta,theta_old) computed in the E-step of EM!!!! this doesn t depend on \theta. ignore it. therefore, EM is optimizing the right thing!